Deep Learning Accelerates the Discovery of Two-Dimensional Catalysts for Hydrogen Evolution Reaction

Abstract

Two-dimensional materials with active sites are expected to replace platinum as large-scale hydrogen production catalysts. However, the rapid discovery of excellent two-dimensional hydrogen evolution reaction catalysts is seriously hindered due to the long experiment cycle and the huge cost of high-throughput calculations of adsorption energies. Considering that the traditional regression models cannot consider all the potential sites on the surface of catalysts, we use a deep learning method with crystal graph convolutional neural networks to accelerate the discovery of high-performance two-dimensional hydrogen evolution reaction catalysts from two-dimensional materials database, with the prediction accuracy as high as 95.2%. The proposed method considers all active sites, screens out 38 high performance catalysts from 6,531 two-dimensional materials, predicts their adsorption energies at different active sites, and determines the potential strongest adsorption sites. The prediction accuracy of the two-dimensional hydrogen evolution reaction catalysts screening strategy proposed in this work is at the density-functional-theory level, but the prediction speed is 10.19 years ahead of the high-throughput screening, demonstrating the capability of crystal graph convolutional neural networks-deep learning method for efficiently discovering high-performance new structures over a wide catalytic materials space.